| 摘要: | 類神經網路在山崩研究的應用多為評估邊坡破壞與否，較少人研究區域性之山崩潛感。Saro Lee（2003）首度運用類神經網路分析預測山崩，再將分析結果合併多變量分析，繪製山崩潛感圖。林彥享（2003）利用類神經網路的自我學習的能力，結合模糊理論，進行廣域性山崩潛感分析之研究並繪製山崩潛感圖。本研究採用林彥享（2003）之分析模式，建立模糊類神經網路分析系統，學習山崩發生的機制和預測可能發生山崩的位置。本研究引入Rprop演算法消除類神經網路訓練時間過長的問題，並將分析結果與其他統計方法作比較，驗證類神經網路於山崩潛感分析上的可行性。 本研究參考地調所山崩潛感分析計畫其研究架構，建立本研究的架構與工作流程。並沿用國姓地區的四次誘發山崩事件（賀伯颱風、集集地震、桃芝颱風與敏督利颱風）資料，包括：岩性、坡向、坡度、地形粗糙度、坡度粗糙度、總曲率、全坡高、NDVI、愛氏震度與最大時雨量等因子資料。山崩目錄由各事件前後之SPOT衛星影像，判釋崩塌地位置，輔以災後像片基本圖及野外查核作檢驗。以亂數對訓練資料做隨機取樣，使用Matlab軟體進行訓練與建立網路架構，再對全區資料進行分析，經由解模糊化將網路輸出轉化為山崩潛感值，並繪製山崩潛感圖。 分析結果顯示，四個事件的高潛感區大都位於坡度較高的地區，和一般認知會發生山崩的地區不謀而合，與專家所圈繪的山崩位置大致上亦十分符合，崩壞比曲線也有潛感值高崩壞比高的趨勢。 線性系統為了不受資料的雜訊影響，常需對資料進行處理。但對類神經網路分析來說，其內部節點間的互相作用，能夠將資料的雜訊去除，得到正確的結果。所以對於山崩潛感來說，類神經網路還是有其存在的必要性，能夠提供傳統統計方法一個參考。 Artificial neural network method had not been applied to regional landslide susceptibility analysis until Lee（2003）which first used this method to evaluate landslide hazard and combine it with the multivariate analysis to construct a landslide susceptibility map. Lin（2003） utilized the artificial neural network and a fuzzy theory to produce a continuous spectrum to indicate landslide susceptibility and use this to draw a landslide susceptibility map. This study follows Lin’s method and tries to refine the fuzzy neural network system in order to learn the mechanism of landslide and to predict the location that a landslide may happen. This study adopts the Rprop algorithm to significantly reduce the long training time in the artificial neural network. Comparing the result with that of two multivariate methods validates that the fuzzy neural network system is suitable for a landslide susceptibility analysis. This study refers to the work scheme of the landslide susceptibility analysis project in the Central Geological Survey, Taiwan（CGS）, to establish the study scheme and work procedure. Following the usage of the factors in the CGS project including lithology, slope, slope aspect, terrain roughness, slope roughness, total curvature, total height, NDVI, Arias intensity and maximum hourly rain fall factors from the four triggering events the Herb typhoon, the ChiChi earthquake, the Toraji typhoon, and the Mindule typhoon, these factors were rechecked. Landslide inventory interpreted from SOPT image, was also checked by examining a series of rectified aerial photographs in GIS and in the field. This study uses random sampling to get the training samples and proceeds to establish the fuzzy neural network framework in Matlab, and then applied the trained network to the whole area. The output fuzzy membership for landslide and nonlandslide was defuzzied, to become a single value indicating landslide susceptibility. These values were used to construct a landslide susceptibility map. The result shows that high susceptibility areas generally locate at high slope areas and fit well with the actual landslide areas that experts interpreted. The landslide ratio has a trend that higher landslide susceptibility index expresses higher landslide ratio for all terrains and for the four events. In order to depress the noise in a linear system, more data processing is usually needed. However, the interaction among the internal nodes in the neural network can clear the noise data and get better result. A fuzzy neural network system is applicable to the landslide susceptibility analysis, and it could provide a reference for comparison with the traditional statistical methods. |
